How to evaluate controlled-environment software
Words like AI, autonomous and open mean very little until a supplier has to show what sits behind them. These questions ask for that, from any vendor, including us.
Why we publish this. A checklist that only we could pass would be marketing. Several questions below are ones we currently answer with a limitation, and those answers are printed with the rest. If a vendor cannot give a specific answer to a specific question, that is the answer.
Architecture and engineering
For the party building or modifying the facility. Software cannot rescue a room that dehumidifies badly.
- Has the provider delivered a farm of this physical scale and crop type before?
- Is airflow modelled at every rack level, not just as a room average?
- Is latent heat and dehumidification load calculated from crop transpiration rather than a rule of thumb?
- Are drainage, cleanability, condensation, corrosion and biosecurity designed in from the start?
- Can racks, LEDs, irrigation and HVAC be serviced without shutting the whole farm down?
- Are electrical loads, harmonics, three-phase distribution and backup modes documented?
Software and data
The questions that separate an operating layer from a vendor dashboard. Our own answers are below.
- Is there a documented API, or only a dashboard?
- Can a third-party system read and write setpoints in real time?
- Which protocols are actually supported: MQTT, Modbus, BACnet, OPC-UA?
- Who owns the sensor data, the images, the crop recipes and any derived models?
- Can everything be exported after the contract ends, and in what format?
- What happens to the site if the internet connection fails?
- What is the control frequency and the command latency?
- Are there hard safety bounds that sit outside the AI model?
- Is the AI predictive, prescriptive or genuinely closed-loop, and which parts are which?
- What independently verified yield, energy or quality improvements exist?
Commercial viability
The questions that decide whether the deployment survives its second year.
- What is the delivered capital cost per square metre of growing area?
- What is the real kWh per kilo for your target crop, measured rather than modelled?
- Are licences charged per site, zone, device, user, crop or kilogram?
- Are commissioning, recipe development, updates and support included or extra?
- Is there a performance guarantee, or only an equipment warranty?
- What happens if the supplier fails, is acquired or discontinues the platform?
What BeeGrow AI says to the software and data list
Every claim below is one you are entitled to test during a trial, and one we expect to be held to in a contract.
Is there a documented API?
Yes. The web and mobile clients are built on the same authenticated REST APIs served through one gateway. External access, key management and webhook event coverage are documented and licensed for the integration endpoints in your solution scope.
Can you read and write setpoints in real time?
Reading, yes, wherever the source publishes. Writing depends on the controller: MQTT-based actuator command is supported, and vendor-specific climate-computer control is a scoped integration validated per make and model, commissioned read-only first.
Which protocols do you support?
MQTT is the primary path for both ingestion and command. Configured Modbus TCP polling is available for supported sensor hardware. Modbus RTU, BACnet, OPC-UA and proprietary buses need a site gateway or an adapter, and we will tell you which before you sign.
Who owns the data?
You do, including sensor history, images, recipes and the operational record. We do not sell customer data. The data processing terms state the processing basis, and contracted providers may process data only to operate the service.
Can everything be exported at the end?
Yes, but not with one button. While you are a customer your records stay available through the web and mobile apps, and through the API where your agreement includes it. At the end of the contract we produce a full copy of your data for you, in the format, timing and retention window written into the agreement. There is no self-service export of a whole account, and we would rather say so now than have you find out when you need it.
What happens if the internet fails?
It depends which connection is lost. A member of staff may still see previously loaded data, but recording something new needs the connection back: there is no offline write queue today. If site readings stop arriving, BeeGrow cannot continue normal feedback control. Equipment behaviour depends on the installed controller, its last command and its local safeguards; automatic handover to a climate computer is not guaranteed. Gateway buffering and the response to connection loss are confirmed and tested for your installation during commissioning.
Are there hard safety bounds outside the model?
Yes. The PID and actuator-test paths use a safety guard with fixed platform bounds. Some commands are blocked; excessive duration or intensity can be capped before sending. Environmental checks depend on the relevant cached sensor reading being available. PID refusals are logged, and successful test-command publication is audited. These checks do not replace local interlocks or commissioning tests.
Is the AI predictive, prescriptive or closed-loop?
The methods differ. Harvest forecasting can use a trained model; growth-stage estimates currently use rules and observations. Tariff comparisons calculate the cost of candidate schedules. PID control adjusts commissioned equipment towards recipe targets. Forecasts and tariff recommendations do not directly control equipment.
What verified improvements can you show?
We do not yet publish independently verified yield, energy or quality improvements. A paid pilot can establish a baseline and measure the change in energy, water, labour, yield and rejection rate for your crops. Any projected saving before that is an estimate.
Baselines worth measuring before anything changes
The only honest way to judge a platform is against your own numbers from before it arrived. Capture these for a full cycle first, whichever supplier you choose.
| Measure | Why it matters |
|---|---|
| kWh per kilo | The dominant cost line in a lit operation, and the one most affected by lighting schedule changes. |
| Daily light integral delivered | Distinguishes a genuine agronomic change from the crop simply receiving less light more cheaply. |
| Water and nutrient per kilo | Catches savings that are really just a shift of loss from one place to another. |
| Labour minutes per kilo | Usually the largest controllable cost after energy, and the least measured. |
| Cycle time | A faster cycle changes annual throughput more than a yield uplift on the same cycle. |
| Yield and rejection rate | Yield alone can rise while saleable output falls. Track both or neither. |
| Gross margin per cycle | The number that decides whether any of the above mattered commercially. |
Put these questions to us
Send the list, or your own version of it, and we will answer in writing before any call. A vendor that will not put an answer in writing has told you something.
- Walkthrough led by someone who has built the system
- We map your zones, crops and sensors before the call
- Straight answers on what fits and what does not